TY - BOOK

T1 - Traffic regulation recognition from GPS data

AU - Zourlidou, Stefania

N1 - Funding Informtaion: I am grateful to the Equal Opportunities Office (Hochschulbüro für ChancenVielfalt) of Leibniz University Hanover for supporting me financially during the last six months of my studies in order to complete and submit my thesis I am also grateful to IAV GmbH for supporting my research for three years, and to the Deutsche Forschungsgemeinschaft (DFG) for also supporting my work at various stages.

PY - 2023

Y1 - 2023

N2 - This dissertation shows the importance of using low-cost crowd-sourced information for the task of traffic regulator recognition (traffic signals, stop signs, priority signs, uncontrolled intersections), the cost of which in terms of time and money is much higher if standard technology is used for surveying. GPS trajectories can reveal the movement patterns of traffic participants, and the initial hypothesis that traffic regulations can be retrieved by mining the movement patterns imposed by traffic rules is verified. The predictive ability of the classifier becomes more accurate when static information derived from open maps (OSM) is merged with dynamic features extracted from GPS trajectories. An extensive evaluation of the proposed methodology on three datasets, provided classification accuracy between 95% and 97%. For recovering incorrect predictions, an additional consistency check of the predicted regu- lation labels based on domain knowledge rules is proposed, which increases the classification accuracy by 1%-3%. The low sampling rate of GPS traces was found that can negatively affect the classification performance, decreasing the classification accuracy between 1%-2% when the sampling interval is doubled from 2 s to 4 s. In contrast, excluding curved trajectories from the analysis has a positive effect on classification performance. It was also shown that the optimal number of trajectories per intersection arm in terms of computing classification features is five straight trajectories. The problem of sparsity of labeled data was investigated by exploring different scenarios of availability of labeled data. Both unsupervised and semi-supervised techniques such as clustering, self-training, cluster-then-label, as well as active learning were examined. The idea of transferability of learning between cities (training a classifier in a city A and predict- ing regulators in a city B) was also tested, discovering the conditions under which it may be feasible and its limitations. The most accurate predictions of the above tested learning methods were achieved through active learning, which was found to reduce the number of labeled data required for training by 66.7% in the two tested datasets. Finally, a hypothetical scenario is described for the first time, to the author’s knowledge, in the field of traffic regulation recognition from GPS data, where information arrives as data streams, opening up further the possibilities to address the traffic regulation recognition problem in an incremental and more dynamic way

AB - This dissertation shows the importance of using low-cost crowd-sourced information for the task of traffic regulator recognition (traffic signals, stop signs, priority signs, uncontrolled intersections), the cost of which in terms of time and money is much higher if standard technology is used for surveying. GPS trajectories can reveal the movement patterns of traffic participants, and the initial hypothesis that traffic regulations can be retrieved by mining the movement patterns imposed by traffic rules is verified. The predictive ability of the classifier becomes more accurate when static information derived from open maps (OSM) is merged with dynamic features extracted from GPS trajectories. An extensive evaluation of the proposed methodology on three datasets, provided classification accuracy between 95% and 97%. For recovering incorrect predictions, an additional consistency check of the predicted regu- lation labels based on domain knowledge rules is proposed, which increases the classification accuracy by 1%-3%. The low sampling rate of GPS traces was found that can negatively affect the classification performance, decreasing the classification accuracy between 1%-2% when the sampling interval is doubled from 2 s to 4 s. In contrast, excluding curved trajectories from the analysis has a positive effect on classification performance. It was also shown that the optimal number of trajectories per intersection arm in terms of computing classification features is five straight trajectories. The problem of sparsity of labeled data was investigated by exploring different scenarios of availability of labeled data. Both unsupervised and semi-supervised techniques such as clustering, self-training, cluster-then-label, as well as active learning were examined. The idea of transferability of learning between cities (training a classifier in a city A and predict- ing regulators in a city B) was also tested, discovering the conditions under which it may be feasible and its limitations. The most accurate predictions of the above tested learning methods were achieved through active learning, which was found to reduce the number of labeled data required for training by 66.7% in the two tested datasets. Finally, a hypothetical scenario is described for the first time, to the author’s knowledge, in the field of traffic regulation recognition from GPS data, where information arrives as data streams, opening up further the possibilities to address the traffic regulation recognition problem in an incremental and more dynamic way

M3 - Doctoral thesis

SN - 978-3-7696-5310-6

T3 - Wissenschaftliche Arbeiten der Fachrichtung Geodäsie und Geoinformatik der Leibniz Universität Hannover

CY - München

ER -